List of Papers Browse by Subject Areas Author List
Abstract
Most advances in whole-slide image (WSI) classification have followed the multiple instance learning (MIL), with multi-resolution (MR) approaches gaining attention under the assumption that higher magnification provides richer evidence. However, high-resolution is not always beneficial: when low magnification already provides sufficient signal, high-magnification details can introduce noise and additional computational cost. This is also inconsistent with clinical practice, where high magnification is typically consulted when the low-magnification overview is inconclusive. Motivated by the coarse-to-fine diagnostic workflow of pathologists, we propose DiagMIL, which reframes MR-WSI analysis for multi-class diagnosis as deciding when to consult high-resolution information. DiagMIL performs low-resolution coarse screening with independent class-wise sigmoid evidence first, and our novel differentiable uncertainty-aware gate selectively activates a high-resolution softmax stage only when coarse evidence is uncertain. To support this decision making, DiagMIL leverages structure-aware graph representations with cross-resolution links and structure-level distillation. Extensive experiments show that DiagMIL consistently outperforms strong baselines. Comprehensive analyses validate our perspective on high-resolution integration, showing that selectively invoking fine-grained evidence is more effective and substantially reduces computational cost. Code is available at https://github.com/zzzzzioni/DiagMIL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4926_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zzzzzioni/DiagMIL
Link to the Dataset(s)
BRACS dataset: https://www.bracs.icar.cnr.it/
BCNB dataset: https://bupt-ai-cz.github.io/BCNB/
BibTex
@InProceedings{JeoJiw_DiagMIL_MICCAI2026,
author = { Jeong, Jiwon AND Hong, Sungrae AND Han, Donghee AND Shin, Jisu AND Kim, Kyungeun AND Yi, Mun Yong},
title = { { DiagMIL: Uncertainty-Gated Coarse-to-Fine Multi-resolution MIL for WSI Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents DiagMIL, an uncertainty-gated coarse-to-fine multi-resolution multiple instance learning (MR-MIL) framework. It is designed to address the noise and computational redundancy caused by the indiscriminate use of all-resolution information in existing MR-MIL methods, while closely mirroring how pathologists actually work in practice. The main contributions of the paper are fourfold: 1.It proposes a two-stage paradigm consisting of Low-resolution Coarse Screening (LrCS) followed by selective High-resolution Fine Diagnosis (HrFD). The high-resolution branch is activated only when the low-resolution evidence is insufficient, with the goal of fundamentally addressing a key weakness of current MR-MIL pipelines. 2.It introduces a differentiable uncertainty-aware gating mechanism. Based on Dirichlet belief modeling of the low-resolution sigmoid outputs, the model combines four uncertainty-related signals—entropy, evidence scarcity, inter-class margin, and positive mass—to make an end-to-end trainable slide-level routing decision. 3.It builds a structure-aware graph representation and a cross-resolution optimization strategy. Tissue structure is modeled using SLIC-K superpixel partitioning together with minimum spanning tree refinement, and cross-resolution links and a structure-level distillation loss are introduced to encourage cross-resolution consistency and improve the robustness of the gating decision. 4.It is validated systematically on one internal 7-class colon WSI dataset and two public multi-class breast WSI datasets (BRACS and BCNB). The method not only outperforms mainstream MR-MIL SOTA baselines on all datasets, but also substantially reduces computational cost and per-slide processing time through selective routing, offering a better balance between performance and practical deployability.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
+The motivation is strong and closely aligned with the pathology workflow: Rather than simply assuming that higher resolution is always better, the paper explicitly argues that when low-resolution evidence is already sufficient, high-resolution details may only add noise and extra cost. High resolution should be consulted only in uncertain cases. This is much closer to how pathologists actually examine slides. +The methodological contribution is clear and reasonably well grounded: The uncertainty-aware gate is not just a hand-crafted thresholding rule. It is built on Dirichlet evidence modeling and supports end-to-end differentiable routing. The structure-aware graph design and cross-resolution distillation are clearly intended to support the same central idea of selective coarse-to-fine diagnosis. Although the paper is somewhat close in spirit to recent uncertainty-focused or pathologist-inspired MR-MIL work (for example, UFC-MIL), the overall design is coherent and not simply a loose combination of existing components. In that sense, it offers a reasonably clear new direction for MR-MIL. +The experimental study is fairly comprehensive: The evaluation includes a large internal multi-class colon dataset and two public breast pathology datasets covering both histologic and molecular subtype classification. The baselines span three groups—classic MIL, multi-resolution MIL, and graph-based MIL—with 11 representative SOTA methods in total, and the single-resolution baselines are all adapted to multi-resolution input, making the comparison broad and reasonably thorough. +The quantitative results are strong: The method reaches SOTA classification performance on all datasets while also reducing FLOPs, per-slide inference time, and high-resolution patch preprocessing cost. This directly addresses one of the main obstacles for bringing WSI models from research settings into clinical use, which is an important point that many MIL papers do not deal with seriously.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
+The probabilistic semantics of the core fusion strategy are not fully coherent, and the methodological justification is not yet sufficient: The low-resolution coarse screening branch (LrCS) uses class-wise independent sigmoid outputs, while the high-resolution fine diagnosis branch (HrFD) uses a multi-class softmax output. However, Eq. (8) linearly mixes these two fundamentally different outputs and then uses the mixed prediction for multi-class cross-entropy supervision. In particular, the sigmoid branch does not define a normalized multi-class distribution, whereas the softmax branch does. Their underlying semantics and statistical assumptions are different, so directly mixing them is not mathematically clean and does not fit naturally with the usual assumptions behind cross-entropy. I understand that the authors want LrCS to behave more like independent evidence rather than final class probabilities, but for exactly that reason, the paper needs to explain much more clearly why the formulation in Eq. (8) is appropriate. +The reproducibility of the internal dataset setting and the fairness of the baseline comparison remain somewhat unclear: All single-resolution MIL baselines are converted by the authors into multi-resolution variants. In addition, the reported BRACS baselines differ substantially from prior literature such as PathTree, especially for ABMIL, DSMIL, TransMIL, and DTFD-MIL, whose ACC and AUC values are generally lower here. Since the current paper uses official splits, MR variants, and a different frozen patch encoder, while PathTree reports 5-fold cross-validation with a unified PLIP encoder, these results may not be directly comparable. The authors should clearly describe the full baseline reproduction protocol and explain why their BRACS baselines differ so much from previously reported numbers, in order to avoid misleading comparisons. +The independent ablation of key components is still insufficient: The paper does not further isolate the contribution of the structure-aware graph construction modules, such as removing SLIC-K superpixel partitioning, MST topology refinement, or cross-resolution links, nor does it isolate the effect of the structure-level distillation loss. As a result, it is still hard to tell which parts of the framework are actually driving the overall performance gain. +There are still a few completeness issues in the manuscript: The conference information for ref [12] is listed as WACV 2025, whereas the public record indicates WACV 2026.In addition, “Intra-Strcuture / Inter-Strcuture” in Fig. 1 contains spelling mistakes.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The framing of the problem is reasonably novel, and I particularly appreciate the effort to reformulate MR-WSI diagnosis as the question of when high-resolution evidence is actually needed. At the same time, although the paper is careful experimentally and shows clear advantages in both performance and practical efficiency, I think the current version still needs to be strengthened in several aspects if it is to be a more secure MICCAI acceptance: 1.further explain the rationale for mixing LrCS sigmoid evidence and HrFD softmax prediction during training; 2.provide more detailed implementation information for all baselines, verify reproduction quality against the official results reported in the original papers, add fair comparisons with the latest SOTA methods on BRACS, and clearly state the full experimental protocol, including dataset splits, feature extraction, and patch extraction, so that the results are easier to trust and reproduce; 3.provide independent ablations for the structure-aware graph module and the structure-level distillation loss, to better support the necessity and advantage of the core design choices; 4.fix the remaining minor writing and formatting issues in the manuscript. Overall, this is a fairly strong WSI-MIL methodology paper with clear novelty and practical value, although there is still some room for improvement.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My main reason for recommending Weak Accept is that the central motivation of the paper is valid, and it is clearly more thoughtful than a straightforward multi-resolution fusion design. In particular, the idea that high-resolution information should not be used unconditionally once the low-resolution evidence is already sufficient is both intuitively appealing and consistent with how pathologists actually work in a coarse-to-fine manner. The framing is therefore quite attractive. At the same time, the quantitative results are strong across three datasets, and the experimental analysis is more complete than what is usually seen in borderline submissions. I believe the paper offers a meaningful new perspective for MR-MIL. The manuscript still has some weaknesses, but I do not see them as fatal. The paper would benefit from a stronger explanation of the mixed-supervision design, clearer and more transparent baseline reproduction details, and additional ablations for some of the key structural modules. Still, these issues do not undermine the main contribution of the work. In my opinion, this is a solid MR-MIL methodology paper with clear novelty and clinical relevance, and it has real acceptance potential for MICCAI. For that reason, I recommend Weak Accept, with the final decision depending on the quality of the rebuttal and the completeness of the authors’ clarifications.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes DiagMIL, a multi-resolution MIL framework for whole-slide image diagnosis that performs selective coarse-to-fine inference. The model first makes low-resolution class-wise predictions, derives uncertainty-related evidence from them, and then uses a differentiable gate to decide whether high-resolution analysis should be invoked. The framework also includes structure-aware graph modelling, cross-resolution links, and structure-level distillation. Empirically, the paper reports improved classification accuracy over several multi-resolution MIL baselines on one in-house colon cohort and two public breast pathology datasets, together with lower end-to-end cost under the evaluation protocol.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a practical issue in computational pathology: high-resolution evidence is expensive, and it is plausible that not every slide requires the same level of magnification. A clinically meaningful motivation.
2.The method is reasonable as an engineering design. The low-resolution coarse screening, uncertainty-aware routing, and high-resolution refinement are aligned with a pathologist-inspired coarse-to-fine workflow, and the graph-based formulation gives the model a structured way to aggregate local and global evidence.
3.The experimental section is relatively comprehensive. The authors include three datasets, multiple multi-resolution MIL baselines, gating ablations, stage comparisons, cost analysis, and qualitative examples.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The main idea is not sufficiently novel relative to prior works. The paper should discuss and compare more carefully against closely related work such as Hong et al., “Diagnose like a real pathologist: An uncertainty-focused approach for trustworthy multi-resolution multiple instance learning” (WACV 2025), “Uncertainty-based Model Acceleration for Cancer Classification in Whole-Slide Images” (BIBM 2022), Yu et al., “Hundredfold Accelerating for Pathological Images Diagnosis and Prognosis through Self-reform Critical Region Focusing” (IJCAI 2024), and Dong et al., “Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning” (CVPR 2025). These works already explore closely related ideas such as uncertainty-aware routing, selective high-resolution processing, or low-resolution pre-screening to avoid unnecessary gigapixel computation.
2.The baseline setting is not strong enough for the current computational pathology field, because the paper does not use pathology foundation model encoders. The method uses a frozen extractor from Caron et al., i.e., DINO-style generic visual pretraining, with 384-dimensional embeddings, but does not compare against modern pathology foundation models (conch, uni, virchow, etc). This is a major issue because on many WSI classification tasks, strong low-magnification pathology foundation model features combined with simple aggregators such as ABMIL or CLAM can already achieve very competitive performance. Without such baselines, it is unclear whether the gain comes from the proposed selective routing mechanism or simply from comparing against outdated feature representations.
3.The interpretability and gating analyses are still limited. The paper provides useful ablations and a few qualitative examples, but it remains unclear how robust the routing behavior is across classes, datasets, and failure cases. In particular, the paper would be stronger if it reported how often the gate escalates by class and where the gating decision fails.
4.The efficiency analysis is still not convincing. In whole-slide pathology, the dominant cost is usually minute-level slide preprocessing, patch extraction, and feature encoding, so small differences in model-side inference time or FLOPs are meaningless. Here, the reported end-to-end saving over ABMIL(MR) is only about one-third of the protocol, whereas a much stronger efficiency baseline would be a simple low-resolution-only pipeline, e.g., ABMIL with only low-magnification patch features, which could be far cheaper. The key unanswered question is therefore not whether DiagMIL is better than always-on MR baselines, but whether invoking high-resolution processing remains necessary at all once strong low-magnification representations are available, especially with pathology foundation models. The paper should include this type of analysis.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a practical problem and reports solid results, but the core idea is insufficiently novel relative to prior selective/uncertainty-guided WSI methods, and the lack of pathology foundation model baselines makes its practical advantage unclear.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The manuscript presents an approach to multi-resolution multiple-instance learning for whole-slide image classification. Against common approaches that consider that high-resolution provides better info, the authors consider that sometimes it is not necessary, even harmful, for the diagnosis. They propose coarse-to-fine workflow, first performing classification from low resolution, and activating the high-resolution analysis only when the model is uncertain. The gating mechanism for the high-resolution activation combines four uncertainty signals to calculate a metric (g) that is thresholded at 0.5 to decide on the need for the high resolution WSI for the final classification. To ensure alignment between both resolutions, they create a graph at each resolution and force structure-level distillation. The methods are tested on an in-house dataset and two public WSI datasets. There is an extensive comparison against other state-of-the art WSI classification methods on the three datasets, showing that the proposed method consistently outperforms the rest with low computational cost. The code is not provided.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The results are robust, there is an extensive comparison: 3 benchmark datasets, comparison with several state of the art methods, a gating component study, thorough analysis of the computational cost. The presented method seems to outperform others at a low computational cost. Avoiding the use of high-resolution WSIs in unnecessary cases is very useful. The authors present an original gating metric to decide when the model is uncertain enough with the low-res to activate the high-res path. The method handles inter and intra-patch interactions and implements structure-level distillation to ensure cross-resolution alignment.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The original contributions of the paper are unclear. The feature encoding and graph construction, and the low resolution (LrCS) and high resolution (HrFD) pipelines are not really presented as novel. Neither is the cross-resolution alignment strategy. The novelty seems to lie in the gating mechanism that decides when to use the high resolution pipeline, with four uncertainty signals (entropy, evidence scarcity, margin and positive mass) combined to create a gating metric that is thresholded at inference. Reproducibility concerns: no code is provided, and no statement about code availability is made. The in-house database is not publicly available either. The qualitative evaluation is limited, shown with a poor figure (Fig. 5). The annotations shown in the figure and compared with the most attended structure given by the method are not described or mentioned. It would have been nice to see a case where a multi-resolution method is confounded by the high res image, while the proposed method does better by using only the low-res image. The choice of the pre-trained encoder to generate patch embeddings is not justified in any way. Moreover, it draws attention that an encoder from a 2021 pre-print is chosen [6], among all the foundation models available in the literature. The training strategy is not completely described. How many epochs were used? What are the three random seeds mentioned used for? Does the model always see the high-res and low-res pair of images during training? The usefulness of the structure-level distillation for cross-resolution consistency is not correctly justified. It adds hyperparameters and complexity to the model, but does it actually improve performance?
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The novelty of the methodological contributions is not clearly described and seems to be limited. There are reproducibility concerns because not the code nor the in-house database are made public. The method description is detailed, but I am not sure it is enough to ensure reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Two reviewers gave rather positive evaluations of the paper with high confidence. However, regarding one weak negtive evaluation, the authors still are encouraged to meticulously address reviewers’ concerns, particularly the confusion about the method novelty, the code availability, the private dataset fairness, and some critical experiment settings (i.e., encoder, training strategy, etc).